C2 - Statistical-dynamical forecasts of tropical rainfall
Other researchers: Eva-Maria Walz (PhD)
Quantitative precipitation forecasting in the tropics is one of the remaining large challenges in numerical weather prediction (NWP) today. Improved forecasts are bound to yield substantial socio-economic benefits, e.g., in water management, energy production, agriculture, and disease prevention. However, our research in Phase 1 has shown that global ensemble prediction systems have low skill in the tropics in general and essentially no skill over northern tropical Africa, even when considering time-space accumulations of several days and hundreds of kilometers. Statistical postprocessing can correct for biases and dispersion errors, but cannot outperform static, climatology-based forecasts over Africa, with unrealistic representations of convective organization being a likely reason. We have demonstrated that a promising avenue for improvement lies in the control on precipitation exerted by tropical (and to a lesser extent extratropical) wave phenomena. Over northern tropical Africa, Kelvin waves and African Easterly Waves explain 10–30% of the local variance of precipitation accumulations on the 3-hourly to daily timescale, whereas the Madden-Julian Oscillation and equatorial Rossby waves dominate variability on longer timescales. Cold and dry intrusions related to extratropical Rossby waves have been identified as another potentially predictable control on tropical rainfall. Initial tests based on spatio-temporal correlation structures, as well as tropical-wave filtering and extrapolation, show promising results for precipitation occurrence.
In Phase 2 we aim to refine and expand these observation-based statistical forecasting methods. In particular, the added value of incorporating NWP information will be examined. We will develop three types of techniques: (a) observation-based full probabilistic quantitative precipitation forecasts for northern tropical Africa; (b) an expansion of (a) to the entire tropics, taking into account seasonal variation and regional peculiarities; and (c) an integration of predictors from global NWP models into the aforementioned statistical models. The latter is particularly timely, as new data resources are becoming available (e.g., Sub-seasonal to Seasonal Prediction (S2S)) and as skill in predicting the coupling between tropical waves and rainfall is improving. We expect the NWP models to add value particularly in areas of zonal gradients or during times with fast temporal changes and/or perturbations from the extratropics. All three developments will require: (a) a careful selection of predictor variables based on meteorological understanding, NWP model skill, and novel ways of automated feature selection tailored to the spatio-temporally correlated setting at hand; (b) statistical model development based on distributional regression concepts, where the forecasts take the form of probability distributions from a parametric family (e.g., Bernoulli-Gamma mixtures); (c) analysis of time-space coherence in the statistical-dynamical forecasts; (d) evaluation of the new techniques against benchmark forecasts; and (e) meteorological interpretation and possibly refinement. We will actively seek collaboration with weather services to work towards an operationalization of our new approaches in order to support decision making and to realize socio-economic benefits.